机械学习课件(日文).pdf
《机械学习课件(日文).pdf》由会员分享,可在线阅读,更多相关《机械学习课件(日文).pdf(80页珍藏版)》请在得力文库 - 分享文档赚钱的网站上搜索。
1、檄械学曾一夕二园I田孝西学院大学情幸艮X亍教育七夕一TUT 2000/06/071一夕7一入力J(D知髡 知的一夕解析.纸才缶亡一儿 大容量亍一夕一入 亍一夕7XV一厅亍 对象(D多棣化Web上(D亍牛入卜L/TUT 2000/06/072知口七入口|抽出|口|奕换J_Jj 0统合外部DBJ 1T/X知TUT 2000/06/0737彳二技法 统tt学 43 5辞戢 g入弓 zL Jly 卜 决定木 Rough set 相 Uli/7b Graph Based Induction J帚系解命理口A 变数逗捐 一夕O)可视化TUT 2000/06/074What is supervised le
2、arning?Input inst ances co nt ains Class at t ribut es Explanat io n at t ribut es.Generat e rules t o describe class descript io ns induct ively.IF co ndit io ns THEN class Learning fro m examples,Inco rpo rat io n o f backgro und kno wledgecf.regressio n,discriminant analysis,neural net wo rk,near
3、est neighbo rTUT 2000/06/075Typical applications Kno wledge acquisit io n t o be used in plant o perat ing expert syst em Act io n predict io n o f o ppo nent t eams in spo rt s mat ch Diagno sis fro m medical t est s Disco very o f act ive mo t ifs in chemical co mpo unds fro m st ruct ure act ivit
4、 y relat io nship dat aset sTUT 2000/06/076Classification of ProblemsTypeOutputUnderstandingExampleClassificationdefinite answers to all questionsUnnecessaryplant operation,character recognitionGuessprobable answers to some questionsUnnecessarysports action prediction,stock price predictionUnderstan
5、dingprobability to all questionsNecessarymedical diagnosis,grammar acquisitionTUT 2000/06/077St reams in learning research I.Classificat io n Pursuit o f Accuracy UCI Repo sit o ry o f machine learning dat abasesMert z,C.J.and Murphy,P.M,(1996):ht t p:/wwwjcs.uci.edu/mleam/MLRepo sit o ry.ht ml St a
6、ndard pro gram fo r co mpariso nQuinlan,J.R,(1993):C4.5:Programs for Machine Learning,Mo rgan Kaufmann;古川(1995):4足上石夕筋只卜勿口.Review 秋集 u a,金田(1998):例力 b o学者1技街必用仁向rrp情辍处理学会Vo l.39,No.2,pp.145-151;No.3,pp.245-251.TUT 2000/06/078决定木方法7眼G色身:鬓G色目的变数青低里 八、青高里 八、茶高里 八、茶高:/口茶低yn y H青低+青高:/口+青高赤+TUT 2000/06/0
7、79决定木青青茶,具具具 枭鼻黑 低嬴高低口八,青:+高,7,口八,,青:+高,7,口八,茶:一 低,7,口八二茶:一TUT 2000/06/0710平均情辍量仁上5变数iiiR-平均情辍量“、p 1 p n 1 nI(p)n)=-log 2-log 2-(p+n)(p+n)(p+n)(p+n)-分前()=-lo g2-lo g2-o o o o=0.954Z;。TUT 2000/06/0711分HI二上盲平均情辍量 利得身;fikcfci)分类真 0.003,/2 :L 均得 高 低 平利2-5 1-33-5 2-3 38-+2-5 1-3 13-5 2-30.97 Ibit0.918bit
8、0.918=0.95 Ibit=0.003bit5 82 g鬓(D色仁上盲分 0.454/;/Y眼(D色仁上分类直0.3476TUT 2000/06/07 12数值属性结合及一儿仁上马糖尿病粉断木SONAR:ht t p:/www.t rl.ibm.co m/pro ject s/s7800/DBmining/index.ht mTUT 2000/06/0713Pro gress in Decisio n Tree Variable with continuous values Entropy gain ratio,Gini index Sampling Pruning Bagging,Boo
9、sting User interface Interactive expansion of a tree Visualization RulesTUT 2000/06/0714G/n/index vs.Ent ro pyGini-index=Z P/(1-Pr)=1-Z P,TUT 2000/06/0715决定木(D方法-秋集金田:例力Jo学雪技街o 必用【二向【十。情辍处理学会Vo L39,No.2,pp.145-151;No.3,pp.245-251(1998).Breiman,L.,Friedman,J.H.,Olshen,R.A.&St o ne,C.J.:Classificat io
10、 n and Regressio n Trees,The Wadswo rt h&Bro o ks/Co le(1984).CART Quinlan,J.R.:C4.5:Pro grams fo r Machine Learning,Mo rgan Kaufmann(1993).古川IR:一夕解析,卜、:(1995).TUT 2000/06/0716St reams in learning research III.Ro ugh set Characteristics Non exploratory Methodology for decision table Analysis of vari
11、able dependencies NP hard to attributes&values References Pawlak,Z.:Ro ugh Set s:Theo ret ical Aspect s o f Reaso ning abo ut Dat a,Kluwer Academic Publishers(1991).W.Ziarko:Review o f Basics o f Ro ugh Set s in t he Co nt ext o f Dat a Mining,Proc.Fourth International Workshop on Rough Sets,Fuzzy S
12、ets,and Machine Discovery,pp.447-457,To kyo(1996).Dat alo gic/R:Reduct Syst ems Inc.TUT 2000/06/0717Ro ugh set rliPositive regionBoundary regionNegative regionTUT 2000/06/0718言十算谩程1:雕散化Obj-112.0218.2319.0 9555 17.5955618.0955719.6955815.7HECT132.217157.075148.0301532.3130175.8182611.260 199.1 1917 4
13、.0 143111.0200117.195186.6422229.9152103.2383241.1161Class SHECT100100210211311100402111512101610100712211800211Reductl=Size,Height,EnergyReduct2=Size,Height,Current Core=Size,HeightTUT 2000/06/0719ClassSHECT100100210211311100402111512101610100712211800211HeightEnergyTemperat ure010021110211221脱明变数目
14、的变数QP=Size,Height,Energy,Current Q=Temperat ureReduct 1(P,Q)=Height,Energy)Reduct 2(P,Q)=Height,Current Co re(P,Q)=Height TUT 2000/06/0720tt算谩程 2:Decisio n mat rix(cjz)62 1 0 2 1 1Rule醇出J123IOBJele3e61e2(S,l)(F;2)(qi)(H0)(E,2)(qi)2)(CD2e4(H2)(C1)(S,0)(ft 2)(Cl)(S,O)(H,2)(C1)3e5(S,1)012)(H2)(H2)4e7(S
15、,1)(H,2)(E,2)(C1)(H,2)(E;2)(C1)(H2)(E,2)(C1)5e8(E,2)(CD(S,0)(H,0)(E,2)(Cl)(S,O)(E,2)(C1)B=BS.1)V(E,2)VC 1)A(QLO)V(E,2)V(C 1)A(E,2)V(C D)=(E,2)V(C,1)B12=(H,2)V(C,1)A(O)V(H,2)V(C,1)A(O)V(H,2)V(C 1)=(H,2)V(C,1)Bl3=(2)A BH,2)A QL2)=(H,2)B14 二(1)V(H,2)V(E,2)V(C,lb A(H,2)V(E,2)V(C,1)A(H,2)V(E,2)V(C 1)=(H,
16、2)V(E,2)V(C,1)B)二(E,2)V(C 1)A(O)V(H,O)V(E,2)V(C,1)A BS,O)V(E,2)V(C,1)=(E,2)V(C,1)(Energy=2)(Current=1)(Height=2)9(Temperat ure=1)(Temperat ure=1)(Temperat ure=1)TUT 2000/06/0721Variable Precisio n Ro ugh Set Mo delPositive regionBoundary regionNegative regionTUT 2000/06/0722Variable Dependency Analy
17、sisNecessary and Sufficient Variable Set sTUT 2000/06/0723Cars exampleNoSizeCylTurboFuelsysDisplace Co mpPo werTransWeightMileage1co mpact6yesEFImedium highhighaut omediummedium2co mpact6noEFImedium mediumhighmanualmediummedium3co mpact4noEFImedium highhighmanualmediummedium4co mpact6yesEFImedium hi
18、ghhighmanuallighthigh5co mpact6noEFImedium medium mediummanualmediummedium6co mpact6no2-BBLmedium medium mediumaut oheavylo w7co mpact6noEFImedium mediumhighmanualheavylo w8subco mpact4no2-BBLsmall highlo wmanuallighthigh9co mpact4no2-BBLsmall highlo wmanualmediummedium10co mpact4no2-BBLsmall highme
19、dium aut o mediummedium11subco mpact4noEFIsmall highlo wmanuallighthigh12subco mpact4noEFImedium medium mediummanualmediumhigh13co mpact4no2-BBLmedium medium mediummanualmediummedium14subco mpact4yesEFIsmall highhighmanualmediumhigh15subco mpact4no2-BBLsmall mediumlo wmanualmediumhigh16co mpact4yesE
20、FImedium mediumhigh highmanual aut omedium mediummedium medium17co mpact6noEFImedium medium18co mpact4noEFImedium mediumhighaut omediummedium19subco mpact4noEFIsmall highmediummanualmediumhigh20co mpact4noEFIsmall highmediummanualmediumhigh21co mpact4no2-BBLsmall highmediummanualmediummediumReduct s
21、(1)cyl,fuelsys,co mp,po wer,weight(2)size,fuelsys,co mp,po wer,weight(3)size,fuelsys,displace,weight(4)size,cyl,fuelsys,po wer,weight(5)cyl,t urbo,fuelsys,displace,co mp,t rans,weight(6)size,cyl,fuelsys,co mp,weight(7)size,cyl,t urbo,fuelsys,t rans,weightCo re:fuelsys,weight Ziarko:The disco very,an
22、alysis,and represent at io n o f dat a dependencies in dat abases,Knowledge Discovery in Databases pp.195-209,Piat et sky-Shapiro&Frawley ed.AAAI Press(1991).TUT 2000/06/0724Reduct&Core Effects to Sum of SquaresSize cyl t urbo fuelsys displace co mp po wer t rans weightVariables 25TUT 2000/06/07Ro u
23、gh Set Met ho d as a To o l o f Dat a Analysis Very go o d rules fo r underst andingDespit e To o many reduct s Number o f reduct s changes wit h co nfidence value in VPRSM Disregard o f frequenciesTUT 2000/06/0726Ro ugh set Pawlak,Z.:Ro ugh Set s:Theo ret ical Aspect s o f Reaso ning abo ut Dat a,K
24、luwer Academic Publishers(1991).W.Ziarko:Review o f Basics o f Ro ugh Set s in t he Co nt ext o f Dat a Mining,Proc.Fourth International Workshop on Rough Sets,Fuzzy Sets,and Machine Discovery,pp.447-457,To kyo(1996).Datalogic/R:Reduct Systems Inc.方法/俞0特徵 雕散表瑰仁对守己方法 共起的玄分布力知得力,可能 tt算量勺一入数kN,属性数占属性值数
25、Uexp(N)TUT 2000/06/07 27St reams in learning research II.Charact erist ic Rules Evaluat io n by Usefulness Pat t erns wit h Accuracy&Suppo rt St at ist ical est imat io n o f generalit y and accuracy 於木(1999):一夕Z一久力、特徵的及一髡兄(D太的(D一般性t正碓性内信赖性同畤FMffi手法、人工知能学会14,139-147.Except io ns as int erest ingness
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- 机械 学习 课件 日文
限制150内